Attention layers are an integral part of modern end-to-end automatic speech recognition systems, for instance as part of the Transformer or Conformer architecture. Attention is typically multi-headed, where each head has an independent set of learned parameters and operates on the same input feature sequence. The output of multi-headed attention is a fusion of the outputs from the individual heads. We empirically analyze the diversity between representations produced by the different attention heads and demonstrate that the heads become highly correlated during the course of training. We investigate a few approaches to increasing attention head diversity, including using different attention mechanisms for each head and auxiliary training loss functions to promote head diversity. We show that introducing diversity-promoting auxiliary loss functions during training is a more effective approach, and obtain WER improvements of up to 6% relative on the Librispeech corpus. Finally, we draw a connection between the diversity of attention heads and the similarity of the gradients of head parameters.
相關內容
Automatic speech recognition (ASR) systems degrade significantly in face of noisy conditions. Recently, speech enhancement (SE) has been introduced as front-end module to reduce noise and improve speech quality for ASR, but it would also suppress some important speech information, i.e., over-suppression problem. To alleviate this, we propose a dual-path style learning approach for end-to-end noise-robust automatic speech recognition (DPSL-ASR). Specifically, we first introduce clean speech feature along with the fused feature from previously proposed IFF-Net as dual-path inputs to recover the over-suppressed information. Then, we propose a style learning method to map the fused feature close to clean feature, in order to learn latent speech information from the latter, i.e., clean "speech style". Furthermore, we employ consistency loss to minimize the distance of ASR outputs in two paths to improve noise-robustness. Experimental results show that the proposed approach achieves relative word error rate (WER) reductions of 10.6% and 8.6% over the best IFF-Net baseline, on RATS Channel-A and CHiME-4 1-Channel Track datasets, respectively. Visualizations of intermediate embeddings indicate that DPSL-ASR can recover abundant over-suppressed information in enhanced speech. Our code is available at GitHub: //github.com/YUCHEN005/DPSL-ASR.
As a unique biometric that can be perceived at a distance, gait has broad applications in person authentication, social security and so on. Existing gait recognition methods pay attention to extracting either spatial or spatiotemporal representations. However, they barely consider extracting diverse motion features, a fundamental characteristic in gaits, from gait sequences. In this paper, we propose a novel motion-aware spatiotemporal feature learning network for gait recognition, termed GaitMAST, which can unleash the potential of motion-aware features. In the shallow layer, specifically, we propose a dual-path frame-level feature extractor, in which one path extracts overall spatiotemporal features and the other extracts motion salient features by focusing on dynamic regions. In the deeper layers, we design a two-branch clip-level feature extractor, in which one focuses on fine-grained spatial information and the other on motion detail preservation. Consequently, our GaitMAST preserves the individual's unique walking patterns well, further enhancing the robustness of spatiotemporal features. Extensive experimental results on two commonly-used cross-view gait datasets demonstrate the superior performance of GaitMAST over existing state-of-the-art methods. On CASIA-B, our model achieves an average rank-1 accuracy of 94.1%. In particular, GaitMAST achieves rank-1 accuracies of 96.1% and 88.1% under the bag-carry and coat wearing conditions, respectively, outperforming the second best by a large margin and demonstrating its robustness against spatial variations.
We propose a novel method that combines CycleGAN and inter-domain losses for semi-supervised end-to-end automatic speech recognition. Inter-domain loss targets the extraction of an intermediate shared representation of speech and text inputs using a shared network. CycleGAN uses cycle-consistent loss and the identity mapping loss to preserve relevant characteristics of the input feature after converting from one domain to another. As such, both approaches are suitable to train end-to-end models on unpaired speech-text inputs. In this paper, we exploit the advantages from both inter-domain loss and CycleGAN to achieve better shared representation of unpaired speech and text inputs and thus improve the speech-to-text mapping. Our experimental results on the WSJ eval92 and Voxforge (non English) show 8~8.5% character error rate reduction over the baseline, and the results on LibriSpeech test_clean also show noticeable improvement.
Unseen noise signal which is not considered in a model training process is difficult to anticipate and would lead to performance degradation. Various methods have been investigated to mitigate unseen noise. In our previous work, an Instance-level Dynamic Filter (IDF) and a Pixel Dynamic Filter (PDF) were proposed to extract noise-robust features. However, the performance of the dynamic filter might be degraded since simple feature pooling is used to reduce the computational resource in the IDF part. In this paper, we propose an efficient dynamic filter to enhance the performance of the dynamic filter. Instead of utilizing the simple feature mean, we separate Time-Frequency (T-F) features as non-overlapping chunks, and separable convolutions are carried out for each feature direction (inter chunks and intra chunks). Additionally, we propose Dynamic Attention Pooling that maps high dimensional features as low dimensional feature embeddings. These methods are applied to the IDF for keyword spotting and speaker verification tasks. We confirm that our proposed method performs better in unseen environments (unseen noise and unseen speakers) than state-of-the-art models.
DNA has immense potential as an emerging data storage medium. The principle of DNA storage is the conversion and flow of digital information between binary code stream, quaternary base, and actual DNA fragments. This process will inevitably introduce errors, posing challenges to accurate data recovery. Sequence reconstruction consists of inferring the DNA reference from a cluster of erroneous copies. A common assumption in existing methods is that all the strands within a cluster are noisy copies originating from the same reference, thereby contributing equally to the reconstruction. However, this is not always valid considering the existence of contaminated sequences caused, for example, by DNA fragmentation and rearrangement during the DNA storage process.This paper proposed a robust multi-read reconstruction model using DNN, which is resilient to contaminated clusters with outlier sequences, as well as to noisy reads with IDS errors. The effectiveness and robustness of the method are validated on three next-generation sequencing datasets, where a series of comparative experiments are performed by simulating varying contamination levels that occurring during the process of DNA storage.
This study addresses robust automatic speech recognition (ASR) by introducing a Conformer-based acoustic model. The proposed model builds on the wide residual bi-directional long short-term memory network (WRBN) with utterance-wise dropout and iterative speaker adaptation, but employs a Conformer encoder instead of the recurrent network. The Conformer encoder uses a convolution-augmented attention mechanism for acoustic modeling. The proposed system is evaluated on the monaural ASR task of the CHiME-4 corpus. Coupled with utterance-wise normalization and speaker adaptation, our model achieves $6.25\%$ word error rate, which outperforms WRBN by $8.4\%$ relatively. In addition, the proposed Conformer-based model is $18.3\%$ smaller in model size and reduces total training time by $79.6\%$.
Few-shot learning aims to learn novel categories from very few samples given some base categories with sufficient training samples. The main challenge of this task is the novel categories are prone to dominated by color, texture, shape of the object or background context (namely specificity), which are distinct for the given few training samples but not common for the corresponding categories (see Figure 1). Fortunately, we find that transferring information of the correlated based categories can help learn the novel concepts and thus avoid the novel concept being dominated by the specificity. Besides, incorporating semantic correlations among different categories can effectively regularize this information transfer. In this work, we represent the semantic correlations in the form of structured knowledge graph and integrate this graph into deep neural networks to promote few-shot learning by a novel Knowledge Graph Transfer Network (KGTN). Specifically, by initializing each node with the classifier weight of the corresponding category, a propagation mechanism is learned to adaptively propagate node message through the graph to explore node interaction and transfer classifier information of the base categories to those of the novel ones. Extensive experiments on the ImageNet dataset show significant performance improvement compared with current leading competitors. Furthermore, we construct an ImageNet-6K dataset that covers larger scale categories, i.e, 6,000 categories, and experiments on this dataset further demonstrate the effectiveness of our proposed model.
Generating texts which express complex ideas spanning multiple sentences requires a structured representation of their content (document plan), but these representations are prohibitively expensive to manually produce. In this work, we address the problem of generating coherent multi-sentence texts from the output of an information extraction system, and in particular a knowledge graph. Graphical knowledge representations are ubiquitous in computing, but pose a significant challenge for text generation techniques due to their non-hierarchical nature, collapsing of long-distance dependencies, and structural variety. We introduce a novel graph transforming encoder which can leverage the relational structure of such knowledge graphs without imposing linearization or hierarchical constraints. Incorporated into an encoder-decoder setup, we provide an end-to-end trainable system for graph-to-text generation that we apply to the domain of scientific text. Automatic and human evaluations show that our technique produces more informative texts which exhibit better document structure than competitive encoder-decoder methods.
We study the problem of named entity recognition (NER) from electronic medical records, which is one of the most fundamental and critical problems for medical text mining. Medical records which are written by clinicians from different specialties usually contain quite different terminologies and writing styles. The difference of specialties and the cost of human annotation makes it particularly difficult to train a universal medical NER system. In this paper, we propose a label-aware double transfer learning framework (La-DTL) for cross-specialty NER, so that a medical NER system designed for one specialty could be conveniently applied to another one with minimal annotation efforts. The transferability is guaranteed by two components: (i) we propose label-aware MMD for feature representation transfer, and (ii) we perform parameter transfer with a theoretical upper bound which is also label aware. We conduct extensive experiments on 12 cross-specialty NER tasks. The experimental results demonstrate that La-DTL provides consistent accuracy improvement over strong baselines. Besides, the promising experimental results on non-medical NER scenarios indicate that La-DTL is potential to be seamlessly adapted to a wide range of NER tasks.
Many natural language processing tasks solely rely on sparse dependencies between a few tokens in a sentence. Soft attention mechanisms show promising performance in modeling local/global dependencies by soft probabilities between every two tokens, but they are not effective and efficient when applied to long sentences. By contrast, hard attention mechanisms directly select a subset of tokens but are difficult and inefficient to train due to their combinatorial nature. In this paper, we integrate both soft and hard attention into one context fusion model, "reinforced self-attention (ReSA)", for the mutual benefit of each other. In ReSA, a hard attention trims a sequence for a soft self-attention to process, while the soft attention feeds reward signals back to facilitate the training of the hard one. For this purpose, we develop a novel hard attention called "reinforced sequence sampling (RSS)", selecting tokens in parallel and trained via policy gradient. Using two RSS modules, ReSA efficiently extracts the sparse dependencies between each pair of selected tokens. We finally propose an RNN/CNN-free sentence-encoding model, "reinforced self-attention network (ReSAN)", solely based on ReSA. It achieves state-of-the-art performance on both Stanford Natural Language Inference (SNLI) and Sentences Involving Compositional Knowledge (SICK) datasets.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191